How much water does AI use?

How much water does one ChatGPT or Gemini prompt use?

Four published estimates, from four kinds of source, put one response anywhere from a fraction of a milliliter to 50 milliliters. They differ mainly in what they count.

OpenAI’s 0.000085 gallons converts to about 0.32 mL at 3,785.41 mL per U.S. gallon (0.000085 × 3,785.41 = 0.322).

Google’s figure leaves out the water used to generate electricity and to train the model. The post that gives OpenAI’s figure does not say what it includes.

Mistral also reported a footprint for its Mistral Large 2 model: 281,000 cubic meters of water (281 million liters), measured as of January 2025 after 18 months of use, which includes upstream impacts such as server manufacturing[3].

Why do the estimates differ so much?

Three choices explain most of the gap.

Which water is counted. A data center consumes water on site, mostly in cooling towers that work by evaporation[5]. Many of the power plants that supply it are cooled the same way, and making chips and servers takes water too[4]. Google counts only the first[1]. The UC Riverside estimate adds the second, which is a large part of why it is so much higher.

Withdrawal or consumption. Withdrawn water is taken in and later returned; consumed water evaporates and does not come back to the local supply[6]. The figures above measure consumption.

Which model, where and when. Newer models and software do more work for each watt. Google says the energy behind a median Gemini text prompt fell 33-fold in the year to May 2025[1]. Place and season matter as much: a cooling tower evaporates more water in hot, dry weather, and the UC Riverside range is wide because it depends on when and where the model is deployed[4].

Why does AI use water at all?

Computer chips give off their electricity as heat, and AI chips use a lot of electricity[7]. Many large data centers move that heat outside with cooling towers, which shed it by evaporating water[5]. The hotter and drier the weather, the more water they use.

Not every data center works that way. Air-cooled and closed-loop liquid-cooled designs use little or no water for cooling[8][9]. Microsoft says data center designs it began in August 2024 consume zero water for cooling and will avoid more than 125 million liters a year per data center[10]. Microsoft said in December 2024 that its sites in Phoenix and Mount Pleasant, Wis., would pilot the design in 2026, with new sites coming online from late 2027[10].

The water used to make the electricity is harder to avoid, because it depends on how the local grid generates power[11].

How much water do U.S. data centers use in total?

The most thorough national estimate is the 2024 United States Data Center Energy Usage Report from Lawrence Berkeley National Laboratory, backed by the Department of Energy. It found that U.S. data centers used 176 terawatt-hours of electricity in 2023, 4.4% of the nation’s total[12], and consumed just over 0.36 liters of water on site for every kilowatt-hour[12]. It expects that rate to rise to between 0.45 and 0.48 liters after 2023[12].

That works out to about 66 billion liters of water consumed on site in 2023[12]. The report puts the water used to generate data centers’ electricity at nearly 800 billion liters[12], roughly 12.1 times as much (800 ÷ 66, using the report’s own rounded totals).

Those are national figures, and water stress is local: the same draw has different effects depending on the supply it comes from.

How much water will a data center near me use?

Usually, the project’s own documents say: water and sewer agreements with the city or utility, county planning packets, state permits and company announcements. Many projects publish no figure at all, and the Census marks those “Not published” rather than guessing.

Tracked projects with the largest stated water use

Active projects only (proposed, approved, under construction or operating). Each figure is as the cited source states it. Open a project for the source and what the figure covers.

See every project’s water figure on Data center water use, or convert any figure into households with the impact calculator.

12,000,000 gal/day is about the daily at-home water use of 40,000 U.S. families, using U.S. EPA WaterSense’s figure of “more than 300 gallons per day” per family (source). Because that figure is a floor, the family count is an upper bound. Many project figures are permit limits or peak estimates, not metered use.

To check a project near you, search by ZIP code on Data centers near me.

Common questions

Does ChatGPT use water?

Yes, indirectly. The data centers that run it are cooled partly by evaporating water, and many of the power plants that supply them use water too. OpenAI’s CEO says an average query uses about 0.000085 gallons, roughly a fifteenth of a teaspoon, but the post gives no method for that number.

Why does AI use so much water?

Per prompt, it doesn’t: the companies’ own figures are a few drops. The totals grow large because of scale. AI chips run hot, many large data centers shed that heat by evaporating water, and AI campuses are big: the Census tracks 47 active projects planned at 1 GW or more.

Is the water lost?

Water that evaporates in a cooling tower is consumed: it leaves the local supply. Water that is withdrawn and returned, as at some power plants, is not consumed.

Do all data centers use water for cooling?

No. Air-cooled and closed-loop designs use little or none for cooling, and Microsoft says its designs since August 2024 use zero water for cooling. A data center’s water use depends on its cooling design and its climate, which is why the Census records each project’s stated figure instead of an average.

How much water does it take to train an AI model?

Few companies say. Mistral reports a footprint of 281,000 cubic meters of water (281 million liters) for its Mistral Large 2 model, measured as of January 2025 after 18 months of use and including upstream impacts such as server manufacturing, in its life-cycle study with Carbone 4 and ADEME. That is not a training-only figure.

Sources

1. Measuring the environmental impact of delivering AI at Google Scale
   Google (arXiv 2508.15734) · 2025-08-21
   4 source excerpts

   “consumes the equivalent of five drops of water (0.26 mL)”

   “the water consumed for cooling machines and associated infrastructure in data centers”

   “active AI accelerator power, host system energy, idle machine capacity, and data center energy overhead”

   “A 33x reduction in per-prompt energy consumption driven by software efficiencies”

2. The Gentle Singularity
   Sam Altman (OpenAI CEO), personal blog · 2025-06-10
   Source excerpt

   “It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.”

3. Our contribution to a global environmental standard for AI
   Mistral AI · 2025-07-22
   2 source excerpts

   “1.14 gCO₂e, 45 mL of water, and 0.16 mg of Sb eq”

   “20,4 ktCO₂e, 281 000 m3 of water consumed, and 660 kg Sb eq”

4. Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models
   Li, Yang, Islam and Ren, UC Riverside (arXiv 2304.03271) · 2023-04
   2 source excerpts

   “a 500ml bottle of water for roughly 10 – 50 medium-length responses”

   “on-site water for data center cooling (scope 1), off-site water for electricity generation (scope 2)”

5. Cooling Water Efficiency Opportunities for Federal Data Centers
   U.S. Department of Energy, Federal Energy Management Program · 2019-01-09
   Source excerpt

   “The condenser water loop carries the heat to the cooling tower where the evaporative process extracts the heat and dissipates it to the surrounding atmosphere.”

6. Water Use Terminology
   U.S. Geological Survey
   2 source excerpts

   “Water removed from the ground or diverted from a surface-water source for use.”

   “The part of water withdrawn that is evaporated, transpired, incorporated into products or crops, consumed by humans or livestock, or otherwise not available for immediate use.”

7. High-Performance Computing Data Center Cooling System Energy Efficiency
   National Renewable Energy Laboratory (nrel.gov) · 2025-04-10
   Source excerpt

   “IT equipment, such as compute clusters and data storage systems, produces heat as a byproduct.”

8. Data Centers and Water Fact Sheet
   Environmental Law Institute · 2026-01
   Source excerpt

   “Air conditioning, fans, and vents used to expel hot air and intake cool air. This can have low water requirements if ambient air temperature is cold (e.g. 'free' cooling), or may require cooling towers.”

9. Sustainable by design: Next-generation datacenters consume zero water for cooling
   Microsoft · 2024-12-09
   Source excerpt

   “These new liquid cooling technologies recycle water through a closed loop. Once the system is filled during construction, it will continually circulate water between the servers and chillers to dissipate heat without requiring a fresh water supply.”

10. Sustainable by design: Next-generation datacenters consume zero water for cooling
    Microsoft · 2024-12-09
    4 source excerpts

    “Starting August 2024, all new Microsoft datacenter designs began using this next-generation cooling technology”

    “avoid the need for more than 125 million liters of water per year per datacenter”

    “Phoenix, Arizona, and Mt. Pleasant, Wisconsin, will pilot zero-water evaporated designs in 2026”

    “These new sites will begin coming online in late 2027.”

11. The environmental footprint of data centers in the United States
    Lawrence Berkeley National Laboratory (Environmental Research Letters) · 2021-05
    Source excerpt

    “Large amounts of water are also required to operate data centers, both directly for liquid cooling and indirectly to produce electricity. ... Our bottom-up approach reveals one-fifth of data center servers direct water footprint comes from moderately to highly water stressed watersheds”

12. 2024 United States Data Center Energy Usage Report
    Lawrence Berkeley National Laboratory (Shehabi et al.) · 2024-12
    5 source excerpts

    “reaching 176 TWh by 2023, representing 4.4% of total U.S. electricity consumption”

    “average WUE, which stays just over 0.36 L/kWh through 2023”

    “reaching between 0.45 and 0.48 L/kWh”

    “By 2023, hyperscale and colocation account for 84% of the 66-billion-liter total”

    “The total indirect water footprint of U.S. data centers is nearly 800 billion liters”

Each source was opened and checked against the excerpt shown. Company statements are identified as such. Figures are as published; where sources differ, each is shown.